dshbase

Blog · Guide

DeepSeek Harness rc.8: multimodal inputs & subagent bundles

August 21, 2026 · dshbase · rewritten field note

Rewritten from Sogou WeChat rc.8 coverage themes plus the public GitHub release notes. Not a verbatim WeChat reprint.

Reference / clue (official release; WeChat mirrors via Sogou)

What installers should notice

rc.8 is the first big public-beta bump after launch week: multimodal inputs, richer subagents, Windows terminal polish, and a pile of gateway/streaming fixes. For a hub view that also tracks catalog gaps, see our rc.8 landing.

Images in the loop

DeepSeek adapters can enable native image requests. /goal and /plan accept image context; the @ menu can pull files and sessions into the task. That does not mean every hosted model path is vision-ready — read vision notes and prefer verified packs when the narrative outruns the API.

Tool-layer “sight” still exists

If the model is text-only, OCR / pixel-stat style tools can still approximate screenshots. Treat that as evidence assembly, not photo understanding. Community packs such as modlens and dsh-vision-toolkit live in that lane.

Claude Code & Codex as bundles

Both can install as Profile Bundles. Codex supports non-interactive permissions and named instances. Details: subagent guide.

Windows PTY

Persistent PowerShell sessions land in the PTY path and are on by default in Minimal — fewer “terminal forgot my env” moments.

Fixes worth knowing

  • Large or many images less likely to blow the request
  • Streaming cancel keeps usable prefixes for the next turn / fork
  • Custom OpenAI-compatible gateways get fewer format/reasoning gaps

dshbase habit

Upgrade the harness, then verify plugins again. Marketing posts move faster than directory checks — use audit before you trust a random vision bridge.

All articles →